Data quality scene inspection method and device, equipment and storage medium

Abstract

The invention relates to the technical field of data processing, and discloses a data quality scene inspection method, device and equipment and a storage medium, initial multi-modal data of different scenes is acquired, and the initial multi-modal data is preprocessed; constructing a situation model library, matching the preprocessed initial multi-modal data with the situation model library, determining whether the preprocessed initial multi-modal data meets situation requirements or not, and obtaining target multi-modal data if the preprocessed initial multi-modal data meets the situation requirements; a generative adversarial network containing multi-task learning is adopted, a shared network structure is utilized to carry out quality evaluation on the target multi-modal data from multiple dimensions, the target multi-modal data is scored on each dimension, and a final quality score is obtained; performing quality optimization on the target multi-modal data by adopting a variational auto-encoder to obtain optimized target multi-modal data; generating a corresponding quality report for the optimized target multi-modal data, and outputting the quality report; according to the method, the performance and the applicability of the data in practical application can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data quality scenario verification method, device, equipment and storage medium. Background Art

[0002] With the continuous development of generative artificial intelligence technology, especially in the fields of natural language processing (NLP), image generation, speech synthesis, etc., generative models have been widely used in content creation, automated writing, advertising production, image design and other fields. Generative AI can generate new data based on training data and specific inputs. These generated data have great advantages in richness and creativity, but they also have problems such as inconsistent quality and errors. Especially when the generated data is applied to actual situations, quality control becomes a challenge. Summary of the invention

[0003] The purpose of the present invention is to solve the above problems and to design a data quality scenario verification method, device, equipment and storage medium.

[0004] A first aspect of the present invention provides a data quality scenario verification method, the data quality scenario verification method comprising:

[0005] Collecting data from various data sources to obtain initial multimodal data corresponding to different scenarios, and completing preprocessing of the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data;

[0006] According to the application field and business needs of the data, different scenarios are defined and corresponding scenario data are generated. A scenario model library is constructed based on the scenario data, and the preprocessed initial multimodal data is matched with the scenario model library to determine whether the preprocessed initial multimodal data meets the scenario requirements. If so, the target multimodal data is obtained.

[0007] A generative adversarial network with multi-task learning is used to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain a quality assessment result, and the target multimodal data is scored in each dimension according to the quality assessment result, and a final quality score is obtained by combining a weighted algorithm;

[0008] When the final quality score is lower than a preset threshold, a variational autoencoder is used to optimize the quality of the target multimodal data to obtain optimized target multimodal data;

[0009] Generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

[0010] Optionally, in a first implementation of the first aspect of the present invention, defining different scenarios and generating corresponding scenario data according to the application field and business requirements of the data, and building a scenario model library based on the scenario data includes:

[0011] According to the application field and business requirements of the data, the K-Means clustering algorithm is used to conduct a preliminary analysis of the obtained data and assign the data points to different clusters, each cluster representing a situation category;

[0012] Based on the situation categories determined by the K-Means clustering algorithm, the decision tree algorithm is used to extract the key features of the clustered data. By continuously splitting nodes, the optimal features and splitting points are selected so that the data in each child node belongs to the same situation category. Different situations are defined according to the structure of the decision tree and the characteristics of the nodes, and the corresponding situation data are generated. The information gain calculation formula of the decision tree is:

[0013]

[0014] In the formula, B represents the data set, A represents the context category, Values(B) represents all possible values ​​of the context category, and B v represents the subset of context categories with value v, Entropy(B) represents the entropy of the data set;

[0015] The situational data is predicted and expanded using a regression analysis algorithm to construct a situational model library M=f(P,Q)→O, where f represents a function process of generating situational data O and constructing a situational model library based on the application domain P and business demand Q.

[0016] Optionally, in a second implementation of the first aspect of the present invention, the obtained data is preliminarily analyzed using a K-Means clustering algorithm according to the application field and business requirements of the data, and the data points are assigned to different clusters, each cluster representing a context category, including:

[0017] Obtain sample data points based on the application field and business requirements of the data to obtain a data set, S. Calculate the density parameter of each sample data point based on the distance between the sample data points, and find the sample point with the largest density as the initial cluster center;

[0018] Calculate the weighted average Euclidean distance of the sample data points and use this as the radius. The points in this area constitute the first cluster set S1:

[0019]

[0020] Where, d wm represents the weighted average Euclidean distance of the sample data points, n represents the number of sample points, dw (x i ,x j ) represents two sample points x i and x j The weighted Euclidean distance between them, m represents the number of features;

[0021] Let S = S-S1, and find the sample point with the largest density parameter in the remaining data As the cluster center, the weighted average Euclidean distance of the sample points is used as the radius to obtain the second cluster set S2, and the iteration is repeated until k cluster centers are found;

[0022] Calculate the distance from each sample data point to the selected k initial centers, divide the sample data points into the nearest clusters to form the initial division, calculate the mean of each class of the initial division as the new center of the class, and calculate the clustering error sum of squares by the error square formula:

[0023]

[0024] In the formula, E represents the sum of squares of clustering errors, k represents the number of clusters, and c j represents the jth cluster, x i ∈c j Represents x i Is c j Centered on;

[0025] Repeat the iteration and compare the sum of squared errors of this division with the sum of squared errors of the previous division. If the absolute value of the difference between the two does not exceed 10 -5 , stop the iteration and output the final clustering result.

[0026] Optionally, in a third implementation of the first aspect of the present invention, matching the preprocessed initial multimodal data with the context model library to determine whether the preprocessed initial multimodal data meets the context requirements, and if so, obtaining the target multimodal data, includes:

[0027] In the process of matching the pre-processed initial multimodal data with the context model library, the SIFT algorithm is used to extract the key features of the image data, and the natural language processing technology is used to convert the text into a numerical vector for the text data;

[0028] After obtaining the key features of the image data and the numerical vectors converted from the text data, the image key feature vector and the text numerical vector are fused by splicing fusion to obtain the fused feature vector F dusion ;

[0029] The cosine similarity algorithm is used to calculate the fusion feature vector F fusionThe similarity between the feature vectors corresponding to each situation model in the situation model library:

[0030]

[0031] In the formula, F situation represents the feature vector corresponding to each situation model in the situation model library, represents the vector dot product operation, ||F fusion || and ‖F situation ‖ respectively represent the modulus of the corresponding vector;

[0032] The calculated fusion feature vector F fusion The data is input into the SVM classifier, and the corresponding category label is output by the SVM classifier to determine whether the preprocessed initial multimodal data meets the situational requirements. The data corresponding to the situational requirements are extracted and output as the target multimodal data, where the category label 0 indicates that the situational requirements are not met, and the category label 1 indicates that the situational requirements are met.

[0033] Optionally, in a fourth implementation of the first aspect of the present invention, the use of a generative adversarial network containing multi-task learning uses a shared network structure to perform quality assessment on the target multimodal data from multiple dimensions to obtain a quality assessment result, including:

[0034] Construct a generative adversarial network with multi-task learning, which consists of a generator and a discriminator. The weights of the discriminator are spectrally normalized. For each layer of the weight matrix W of the discriminator, the spectral norm of the weight matrix W is calculated:

[0035]

[0036] Divide the weight matrix W by the spectral norm:

[0037]

[0038] Where ‖Wh‖2 represents the Euclidean norm of the vector Wh, W sn represents the weight matrix after spectral normalization;

[0039] In the generative adversarial network, the mean square error loss function is used to evaluate the image task, and the cross entropy loss function is used to evaluate the text grammatical correctness task. The total loss function is:

[0040]

[0041] Where, L GAN (G,D) represents the loss function of the generative adversarial network. Multi-task loss function, λ i Represents the weight parameter that balances the losses of different tasks;

[0042] The target multimodal data is input into the generative adversarial network, and different output branches of the generative adversarial network correspond to quality assessment results of different dimensions.

[0043] Optionally, in a fifth implementation manner of the first aspect of the present invention, scoring the target multimodal data in each dimension according to the quality assessment result and obtaining a final quality score in combination with a weighted algorithm includes:

[0044] Assume that the target multimodal data has n dimensions, namely D1, D2, …, D n , the target multimodal data is in dimension D i The score on Calculate each dimension D of the target multimodal data i Information entropy E i :

[0045]

[0046] Calculate each dimension D i The weight w i :

[0047]

[0048] In the formula, m represents the number of data samples, p ij Represents the probability of the jth sample in the i-th dimension;

[0049] According to the weighted algorithm and the obtained weights, the final quality score of the target multimodal data is calculated.

[0050] Optionally, in a sixth implementation manner of the first aspect of the present invention, when the final quality score is lower than a preset threshold, using a variational autoencoder to optimize the quality of the target multimodal data to obtain optimized target multimodal data, including:

[0051] Construct a variational autoencoder, which consists of an encoder and a decoder, input low-quality target multimodal data into the encoder, and the encoder maps the target multimodal data to a latent space that follows a normal distribution to obtain latent variables, adjust the latent variables, explore different areas in the latent space, find better latent representations, and obtain adjusted latent variables, wherein the latent variable adjustment method at least includes adjusting tone, modifying grammatical structure, and optimizing image details;

[0052] The adjusted latent variables are input into the decoder for reconstruction to obtain the optimized target multimodal data, the quality of the optimized target multimodal data is evaluated, and the final quality score is calculated again.

[0053] A second aspect of the present invention provides a data quality scenario verification device, the data quality scenario verification device comprising:

[0054] A data preprocessing module is used to collect data from various data sources, obtain initial multimodal data corresponding to different scenarios, and complete preprocessing of the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data;

[0055] A context matching module is used to define different contexts and generate corresponding context data according to the application field and business requirements of the data, build a context model library based on the context data, and match the preprocessed initial multimodal data with the context model library to determine whether the preprocessed initial multimodal data meets the context requirements. If so, the target multimodal data is obtained;

[0056] A multi-dimensional quality assessment module, which is used to use a generative adversarial network with multi-task learning to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain a quality assessment result, and to score the target multimodal data in each dimension according to the quality assessment result, and to obtain a final quality score in combination with a weighted algorithm;

[0057] A feedback and adaptive optimization module, used for optimizing the quality of the target multimodal data using a variational autoencoder when the final quality score is lower than a preset threshold, to obtain optimized target multimodal data;

[0058] The result output module is used to generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

[0059] The third aspect of the present invention provides a data quality scenario verification device, which includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the data quality scenario verification device performs each step of the data quality scenario verification method described in any one of the above items.

[0060] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, and when the instructions are executed by a processor, the steps of the data quality scenario verification method as described in any one of the above items are implemented.

[0061] In the technical solution provided by the present invention, data is collected from various data sources to obtain initial multimodal data corresponding to different scenarios, and the initial multimodal data is preprocessed through standardization and denoising to obtain preprocessed initial multimodal data; different scenarios are defined and corresponding scenario data are generated according to the application field and business needs of the data, a scenario model library is constructed based on the scenario data, and the preprocessed initial multimodal data is matched with the scenario model library to determine whether the preprocessed initial multimodal data meets the scenario requirements, and if so, the target multimodal data is obtained; a generative adversarial network containing multi-task learning is used to generate the target multimodal data using a shared network structure. Quality assessment is performed from multiple dimensions to obtain quality assessment results, and the target multimodal data is scored in each dimension according to the quality assessment results, and a final quality score is obtained by combining a weighted algorithm; when the final quality score is lower than a preset threshold, a variational autoencoder is used to perform quality optimization on the target multimodal data to obtain optimized target multimodal data; a corresponding quality report is generated for the optimized target multimodal data, and the optimized target multimodal data and the quality report are output; the present invention solves the shortcomings of the prior art in multi-dimensional and multi-scenario quality inspection through contextualized quality assessment and adaptive optimization mechanism, and can effectively improve the performance and applicability of generated data in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0063] Figure 1 A schematic diagram of a first embodiment of a data quality scenario verification method provided by an embodiment of the present invention;

[0064] Figure 2 A schematic diagram of a second embodiment of the data quality scenario verification method provided by an embodiment of the present invention;

[0065] Figure 3 A schematic diagram of a third embodiment of the data quality scenario verification method provided by an embodiment of the present invention;

[0066] Figure 4 A schematic diagram of the structure of a data quality scenario verification device provided by an embodiment of the present invention;

[0067] Figure 5 A schematic diagram of the structure of a data quality scenario verification device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A first embodiment schematic diagram of a data quality scenario verification method provided by an embodiment of the present invention, the method specifically comprises the following steps:

[0070] Step 101: Collect data from various data sources to obtain initial multimodal data corresponding to different scenarios, and preprocess the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data;

[0071] In this embodiment, in the data collection stage: data sources may include sensor networks, social media platforms, database systems, image and video acquisition devices, etc. The initial multimodal data corresponding to different scenarios are diverse. For example, in smart traffic scenarios, they may include vehicle location information (GPS data), speed data, image data (images taken by road cameras), and sound data (vehicle horn sounds, environmental noise, etc.).

[0072] Step 102: define different scenarios and generate corresponding scenario data according to the application field and business requirements of the data, build a scenario model library based on the scenario data, match the preprocessed initial multimodal data with the scenario model library, and determine whether the preprocessed initial multimodal data meets the scenario requirements. If so, obtain the target multimodal data;

[0073] In this embodiment, different scenarios are defined according to the application field and business needs of the data. For example, in the field of intelligent transportation, "traffic congestion scenario", "traffic accident scenario", "normal driving scenario", etc. can be defined. The definition of the scenario can be based on multiple factors, such as time, location, event type, user behavior, etc. For example, in the traffic congestion scenario, factors such as time (peak hours), location (urban main roads), event type (vehicle queue length exceeds a certain threshold) can be considered. For each scenario, its characteristics and attributes can be further refined. For example, in the traffic accident scenario, attributes such as accident type (collision, rear-end collision, rollover, etc.), accident severity (minor, medium, severe), accident location (highway, urban road, rural road, etc.) can be defined.

[0074] Step 103: adopt a generative adversarial network with multi-task learning, use a shared network structure to perform quality assessment on the target multimodal data from multiple dimensions, obtain a quality assessment result, and score the target multimodal data in each dimension according to the quality assessment result, and obtain a final quality score in combination with a weighted algorithm;

[0075] Step 104: When the final quality score is lower than a preset threshold, a variational autoencoder is used to optimize the quality of the target multimodal data to obtain optimized target multimodal data;

[0076] In this embodiment, a generative adversarial network with multi-task learning is constructed. The generative adversarial network consists of a generator and a discriminator. The weight of the discriminator is spectrally normalized. For each layer of the weight matrix W of the discriminator, the spectral norm of the weight matrix W is calculated:

[0077]

[0078] Divide the weight matrix W by the spectral norm:

[0079]

[0080] Where ‖Wh‖2 represents the Euclidean norm of the vector Wh, W sn represents the weight matrix after spectral normalization;

[0081] In the generative adversarial network, the mean square error loss function is used to evaluate the image task, and the cross entropy loss function is used to evaluate the text grammatical correctness task. The total loss function is:

[0082]

[0083] Where, L GAN (G,D) represents the loss function of the generative adversarial network. Multi-task loss function, λ i Represents the weight parameter that balances the losses of different tasks;

[0084] The target multimodal data is input into the generative adversarial network, and different output branches of the generative adversarial network correspond to quality assessment results of different dimensions.

[0085] In this embodiment, it is assumed that the target multimodal data has n dimensions, namely D1, D2, ..., D n , the target multimodal data is in dimension D i The score on Calculate each dimension D of the target multimodal data i Information entropy E i :

[0086]

[0087] Calculate each dimension D i The weight w i , according to the weighted algorithm and the obtained weights, the final quality score of the target multimodal data is calculated:

[0088]

[0089] In the formula, m represents the number of data samples, p ij Represents the probability of the jth sample in the i-th dimension.

[0090] Step 105: Generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

[0091] In this embodiment, a variational autoencoder is constructed, which consists of an encoder and a decoder. Low-quality target multimodal data is input into the encoder, and the encoder maps the target multimodal data to a latent space that follows a normal distribution to obtain latent variables. The latent variables are adjusted to explore different areas in the latent space to find better latent representations to obtain adjusted latent variables, wherein the latent variable adjustment method at least includes adjusting tone, modifying grammatical structure, and optimizing image details; the adjusted latent variables are input into the decoder for reconstruction to obtain optimized target multimodal data, the optimized target multimodal data is quality evaluated, and the final quality score is calculated again.

[0092] In the embodiment of the present invention, by comparing the multimodal data with the predefined context model library, the generated data can be dynamically adjusted according to different application scenarios. This context adaptability makes the generated data not only grammatically and semantically correct, but also in line with actual needs in terms of emotion, style, etc. For example, in dialogue generation, the generated text can automatically adjust the emotional tendency according to the user's emotions, which is in line with the natural fluency of the dialogue; the multi-dimensional quality assessment framework can not only evaluate a single dimension (such as grammar or image clarity), but also comprehensively consider the quality of multiple levels such as emotional expression, style matching, and logical consistency. Through the weighted algorithm, the quality can be comprehensively evaluated according to the needs of different tasks, so as to obtain a comprehensive and accurate quality score; through automatic learning and optimization, the quality of the generated data can be gradually improved in actual applications to avoid low-quality data that may appear in a single generation; the method can adapt to the needs of different generation tasks, such as text generation, image generation, audio generation, etc., and can adjust the context model according to the new application scenario to ensure the wide applicability of the generated data in various tasks.

[0093] See also Figure 2 , a schematic diagram of a second embodiment of a data quality scenario verification method provided by an embodiment of the present invention, the method comprising:

[0094] Step 201: Based on the application field and business requirements of the data, a K-Means clustering algorithm is used to perform a preliminary analysis on the obtained data, and the data points are assigned to different clusters, each cluster representing a situation category;

[0095] In this embodiment, the number of clusters is determined according to the application field and business requirements of the data. For example, in the field of intelligent transportation, if the traffic conditions are to be divided into three categories: congested, smooth, and general, for traffic data, features such as vehicle speed, traffic volume, and road occupancy can be selected. The data points in each cluster have similar features and represent a specific scenario. For example, in traffic data, one cluster may represent a congested scenario with heavy traffic volume and slow speed; another cluster may represent a smooth scenario with moderate traffic volume and fast speed.

[0096] Step 202: Based on the situation categories determined by the K-Means clustering algorithm, the decision tree algorithm is used to extract key features from the clustered data, and the optimal features and split points are selected by continuously splitting nodes so that the data in each child node belongs to the same situation category. Different situations are defined according to the structure of the decision tree and the characteristics of the nodes, and corresponding situation data are generated;

[0097] In this embodiment, the information gain calculation formula of the decision tree is:

[0098]

[0099] In the formula, B represents the data set, A represents the context category, Values(B) represents all possible values ​​of the context category, and B v represents the subset of context categories with value v, Entropy(B) represents the entropy of the data set;

[0100] In this embodiment, sample data points are obtained according to the application field and business requirements of the data to obtain a data set, S. The density parameter of each sample data point is calculated according to the distance of the sample data points, and the sample point with the largest density is found as the initial cluster center; the weighted average Euclidean distance of the sample data points is calculated, and this is used as the radius. The points located in this area constitute the first cluster set S1:

[0101]

[0102] Where, d wm represents the weighted average Euclidean distance of the sample data points, n represents the number of sample points, d w (x i ,x j ) represents two sample points x i and j The weighted Euclidean distance between them, m represents the number of features; let S = S-S1, find the sample point with the largest density parameter in the remaining data As the cluster center, with the weighted average Euclidean distance of the sample points as the radius, the second cluster set S2 is obtained, and the iteration is repeated until k cluster centers are found; the distance from each sample data point to the selected k initial centers is calculated, and the sample data points are divided into the nearest clusters to form an initial division. The mean of each class of the initial division is calculated as the new center of the class, and the sum of squares of clustering errors is calculated by the square error formula:

[0103]

[0104] In the formula, E represents the sum of squares of clustering errors, k represents the number of clusters, and c j represents the jth cluster, x i ∈c j Represents x i Is c j Repeat the iteration and compare the sum of squares of the errors of this division with the sum of squares of the errors of the previous division. If the absolute value of the difference between the two does not exceed 10 -5 , stop the iteration and output the final clustering result.

[0105] Step 203: predicting and expanding the situational data using a regression analysis algorithm to construct a situational model library;

[0106] In this embodiment, the situation model library M = f(P, Q) → O, where f represents the function process of generating situation data O and constructing the situation model library according to the application field P and the business demand Q; select appropriate independent variables and dependent variables. For situation data, some situation-related features can be selected as independent variables, such as time, location, event type, etc., and one or more situation-related indicators can be selected as dependent variables, such as traffic flow, temperature, humidity, etc., to establish a regression model. According to the selected independent variables and dependent variables, a suitable regression analysis algorithm can be selected to establish the model. For example, if there is a linear relationship between the dependent variable and the independent variable, a linear regression algorithm can be selected; if there is a nonlinear relationship, a polynomial regression algorithm can be selected; the established regression model is used to predict the situation data. The new independent variable value is input into the regression model to obtain the predicted value of the dependent variable. For example, in traffic flow prediction, the traffic flow at a certain moment in the future can be predicted based on factors such as time and location.

[0107] See also Figure 3 , a schematic diagram of a third embodiment of a data quality scenario verification method provided by an embodiment of the present invention, the method comprising:

[0108] Step 301: In the process of matching the pre-processed initial multimodal data with the context model library, the SIFT algorithm is used to extract the key features of the image data, and the natural language processing technology is used to convert the text data into a numerical vector;

[0109] In this embodiment, the SIFT algorithm constructs a scale space of the image and detects stable key points at different scales. These key points are usually local extreme points in the image and have a certain invariance to the scale change and rotation of the image. For each detected key point, the gradient direction and amplitude of the surrounding area are calculated, and this information is encoded into a high-dimensional descriptor. The extracted image key features can be used to compare with the image features in the situation model library to determine whether the image data meets specific situation requirements. For example, in a traffic monitoring scenario, key features of road images, such as road signs, vehicle shapes, etc., can be extracted and matched with model libraries under different traffic situations to determine which situation the current traffic condition belongs to.

[0110] Step 302: After respectively acquiring the key features of the image data and the numerical vectors after the text data is converted, the key feature vectors of the image and the numerical vectors of the text are fused by splicing and fusion to obtain a fused feature vector;

[0111] Step 303: Calculate the similarity between the fused feature vector and the feature vectors corresponding to each context model in the context model library using a cosine similarity algorithm;

[0112] In this embodiment, the similarity calculation formula is:

[0113]

[0114] In the formula, F situation represents the feature vector corresponding to each situation model in the situation model library, represents the vector dot product operation, ||F fusion || and ‖F situation ‖ respectively represent the modulus of the corresponding vector, F fusion Represents the fused feature vector;

[0115] Step 304: input the calculated fusion feature vector into the SVM classifier, use the SVM classifier to output the corresponding category label to determine whether the preprocessed initial multimodal data meets the context requirements, and extract and output the data that meets the context requirements as the target multimodal data.

[0116] In this embodiment, the category label 0 indicates that the situation requirement is not met, and the category label 1 indicates that the situation requirement is met.

[0117] See also Figure 4 , a schematic diagram of the structure of a data quality scenario verification device provided by an embodiment of the present invention, the device comprises:

[0118] The data preprocessing module 401 is used to collect data from various data sources, obtain initial multimodal data corresponding to different scenarios, and complete preprocessing of the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data;

[0119] A context matching module 402 is used to define different contexts and generate corresponding context data according to the application field and business requirements of the data, build a context model library based on the context data, and match the preprocessed initial multimodal data with the context model library to determine whether the preprocessed initial multimodal data meets the context requirements. If so, the target multimodal data is obtained;

[0120] The multi-dimensional quality assessment module 403 is used to use a generative adversarial network with multi-task learning to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain a quality assessment result, and to score the target multimodal data in each dimension according to the quality assessment result, and to obtain a final quality score in combination with a weighted algorithm;

[0121] Feedback and adaptive optimization module 404, used to optimize the quality of target multimodal data using a variational autoencoder to obtain optimized target multimodal data when the final quality score is lower than a preset threshold;

[0122] The result output module 405 is used to generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

[0123] In the embodiment of the present invention, by establishing contextualized quality standards, the applicability and effectiveness of generated data in specific application tasks are comprehensively considered, and based on a multi-dimensional evaluation framework, the quality of the data is comprehensively checked, mainly evaluating the contextual adaptability, emotional consistency, scene consistency and logical coherence of the data, and dynamically self-optimizing according to task requirements and application scenarios to ensure high-quality performance of data in multiple scenarios; through automated quality inspection processes and adaptive feedback mechanisms, real-time evaluation and optimization of generated data can be achieved, with high flexibility and intelligence, and can adapt to different fields and application scenarios, such as dialogue generation, image generation, audio generation, etc. Compared with the prior art, the solution of the present invention not only improves the quality of the data, but also enhances the application effect in multiple tasks and multiple scenarios, and has strong generalization ability and practical value. Through contextualized quality evaluation and adaptive optimization mechanisms, the deficiencies of the prior art in multi-dimensional and multi-scenario quality inspection are solved, and the performance and applicability of generated data in practical applications can be effectively improved.

[0124] above Figure 4 The data quality scenario verification device in the embodiment of the present invention is described in detail from the perspective of modular functional entities, and the data quality scenario verification device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0125] Figure 5 It is a structural diagram of a data quality scenario verification device provided by an embodiment of the present invention. The data quality scenario verification device 500 may have relatively large differences due to different configurations or performances, and may include one or more processors 510 (for example, one or more processors) and a memory 520, and one or more storage media 530 storing application programs 533 or data 532. Among them, the memory 520 and the storage medium 530 may be temporary storage or permanent storage. The program stored in the storage medium 530 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the data quality scenario verification device 500. Furthermore, the processor 510 may be configured to communicate with the storage medium 530 to execute a series of instruction operations in the storage medium 530 on the data quality scenario verification device 500.

[0126] The data quality scenario verification device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input and output interfaces 560, and / or one or more operating systems 531, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 5 The structure of the data quality scenario verification device shown does not constitute a limitation on the data quality scenario verification device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0127] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes each step of the data quality scenario verification method provided in the above embodiments.

[0128] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described equipment, devices, and units can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0129] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0130] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data quality scenario inspection method, characterized in that: The data quality scenario verification method includes: Collecting data from various data sources to obtain initial multimodal data corresponding to different scenarios, and completing preprocessing of the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data; According to the application field and business needs of the data, different scenarios are defined and corresponding scenario data are generated. A scenario model library is constructed based on the scenario data, and the preprocessed initial multimodal data is matched with the scenario model library to determine whether the preprocessed initial multimodal data meets the scenario requirements. If so, the target multimodal data is obtained. A generative adversarial network with multi-task learning is used to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain a quality assessment result, and the target multimodal data is scored in each dimension according to the quality assessment result, and a final quality score is obtained by combining a weighted algorithm; When the final quality score is lower than a preset threshold, a variational autoencoder is used to optimize the quality of the target multimodal data to obtain optimized target multimodal data; Generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

2. A data quality scenario verification method according to claim 1, characterized in that: Defining different scenarios and generating corresponding scenario data according to the application fields and business requirements of the data, and building a scenario model library based on the scenario data, includes: According to the application field and business requirements of the data, the K-Means clustering algorithm is used to conduct a preliminary analysis of the obtained data and assign the data points to different clusters, each cluster representing a situation category; Based on the situation categories determined by the K-Means clustering algorithm, the decision tree algorithm is used to extract the key features of the clustered data. By continuously splitting nodes, the optimal features and splitting points are selected so that the data in each child node belongs to the same situation category. Different situations are defined according to the structure of the decision tree and the characteristics of the nodes, and the corresponding situation data are generated. The information gain calculation formula of the decision tree is: In the formula, B represents the data set, A represents the context category, Values(B) represents all possible values ​​of the context category, and B v represents the subset of context categories with value v, Entropy(B) represents the entropy of the data set; The situational data is predicted and expanded using a regression analysis algorithm to construct a situational model library M=f(P, Q)→O, where f represents a function process of generating situational data O and constructing a situational model library based on the application domain P and business demand Q.

3. A data quality scenario verification method as claimed in claim 2, characterized in that: According to the application field and business needs of the data, the K-Means clustering algorithm is used to perform a preliminary analysis of the obtained data, and the data points are assigned to different clusters. Each cluster represents a situation category, including: Obtain sample data points based on the application field and business requirements of the data to obtain a data set, S. Calculate the density parameter of each sample data point based on the distance between the sample data points, and find the sample point with the largest density as the initial cluster center; Calculate the weighted average Euclidean distance of the sample data points and use this as the radius. The points in this area constitute the first cluster set S1: Where, d wm represents the weighted average Euclidean distance of the sample data points, n represents the number of sample points, d w (x i , x j ) represents two sample points x i and x j The weighted Euclidean distance between them, m represents the number of features; Let S = S-S1, and find the sample point with the largest density parameter in the remaining data As the cluster center, the weighted average Euclidean distance of the sample points is used as the radius to obtain the second cluster set S2, and the iteration is repeated until k cluster centers are found; Calculate the distance from each sample data point to the selected k initial centers, divide the sample data points into the nearest clusters to form the initial division, calculate the mean of each class of the initial division as the new center of the class, and calculate the clustering error sum of squares by the error square formula: In the formula, E represents the sum of squares of clustering errors, k represents the number of clusters, and c j represents the jth cluster, x i ∈c j Represents x i It is c j Centered on; Repeat the iteration and compare the sum of squared errors of this division with the sum of squared errors of the previous division. If the absolute value of the difference between the two does not exceed 10 -5 , stop the iteration and output the final clustering result.

4. A data quality scenario verification method according to claim 1, characterized in that: The preprocessed initial multimodal data is matched with the context model library to determine whether the preprocessed initial multimodal data meets the context requirements, and if so, the target multimodal data is obtained, including: In the process of matching the pre-processed initial multimodal data with the context model library, the SIFT algorithm is used to extract the key features of the image data, and the natural language processing technology is used to convert the text into a numerical vector for the text data; After obtaining the key features of the image data and the numerical vectors converted from the text data, the image key feature vector and the text numerical vector are fused by splicing fusion to obtain the fused feature vector F fusion ; The cosine similarity algorithm is used to calculate the fusion feature vector F fusion The similarity between the feature vectors corresponding to each situation model in the situation model library: In the formula, F situation represents the feature vector corresponding to each situation model in the situation model library, represents the vector dot product operation, ||F fusion || and ||F situation || respectively represent the modulus of the corresponding vector; The calculated fusion feature vector F fusion The data is input into the SVM classifier, and the corresponding category label is output by the SVM classifier to determine whether the preprocessed initial multimodal data meets the situational requirements. The data corresponding to the situational requirements are extracted and output as the target multimodal data, where the category label 0 indicates that the situational requirements are not met, and the category label 1 indicates that the situational requirements are met.

5. A data quality scenario verification method according to claim 1, characterized in that: The generative adversarial network with multi-task learning is used to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain quality assessment results, including: Construct a generative adversarial network with multi-task learning, which consists of a generator and a discriminator. The weights of the discriminator are spectrally normalized. For each layer of the weight matrix W of the discriminator, the spectral norm of the weight matrix W is calculated: Divide the weight matrix W by the spectral norm: Where ||Wh||2 represents the Euclidean norm of the vector Wh, W sn Represents the weight matrix after spectral normalization: In the generative adversarial network, the mean square error loss function is used to evaluate the image task, and the cross entropy loss function is used to evaluate the text grammatical correctness task. The total loss function is: Where, L GAN (G, D) represents the loss function of the generative adversarial network. Multi-task loss function, λ i Represents the weight parameter that balances the losses of different tasks; The target multimodal data is input into the generative adversarial network, and different output branches of the generative adversarial network correspond to quality assessment results of different dimensions.

6. A data quality scenario verification method according to claim 1, characterized in that: Scoring the target multimodal data in each dimension according to the quality assessment result and obtaining a final quality score in combination with a weighted algorithm include: Assume that the target multimodal data has n dimensions, namely D1, D2, ..., D n , the target multimodal data is in dimension D i The score on Calculate each dimension D of the target multimodal data i Information entropy E i : Calculate each dimension D i The weight w i : In the formula, m represents the number of data samples, p ij Represents the probability of the jth sample in the i-th dimension; According to the weighted algorithm and the obtained weights, the final quality score of the target multimodal data is calculated.

7. A data quality scenario verification method according to claim 1, characterized in that: When the final quality score is lower than a preset threshold, a variational autoencoder is used to optimize the quality of the target multimodal data to obtain optimized target multimodal data, including: Construct a variational autoencoder, which consists of an encoder and a decoder, input low-quality target multimodal data into the encoder, and the encoder maps the target multimodal data to a latent space that follows a normal distribution to obtain latent variables, adjust the latent variables, explore different areas in the latent space, find better latent representations, and obtain adjusted latent variables, wherein the latent variable adjustment method at least includes adjusting tone, modifying grammatical structure, and optimizing image details; The adjusted latent variables are input into the decoder for reconstruction to obtain the optimized target multimodal data, the quality of the optimized target multimodal data is evaluated, and the final quality score is calculated again.

8. A data quality scenario verification device, characterized in that: The data quality scenario inspection device comprises: A data preprocessing module is used to collect data from various data sources, obtain initial multimodal data corresponding to different scenarios, and complete preprocessing of the initial multimodal data through standardization and denoising to obtain preprocessed initial multimodal data; A context matching module is used to define different contexts and generate corresponding context data according to the application field and business requirements of the data, build a context model library based on the context data, and match the preprocessed initial multimodal data with the context model library to determine whether the preprocessed initial multimodal data meets the context requirements. If so, the target multimodal data is obtained; A multi-dimensional quality assessment module, which is used to use a generative adversarial network with multi-task learning to perform quality assessment on the target multimodal data from multiple dimensions using a shared network structure to obtain a quality assessment result, and to score the target multimodal data in each dimension according to the quality assessment result, and to obtain a final quality score in combination with a weighted algorithm; A feedback and adaptive optimization module, used for optimizing the quality of the target multimodal data using a variational autoencoder when the final quality score is lower than a preset threshold, to obtain optimized target multimodal data; The result output module is used to generate a corresponding quality report for the optimized target multimodal data, and output the optimized target multimodal data and the quality report.

9. A data quality scenario inspection device, characterized in that: The data quality scenario verification device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the data quality scenario verification device executes each step of the data quality scenario verification method as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the various steps of the data quality scenario verification method according to any one of claims 1 to 7 are implemented.

Images